We have been seeing for years how Deepfakes are capable of making many people believe that the subject they are seeing on the screen is, in fact, a real human. The accuracy of AI-generated faces is increasing, and it's no longer just making them hard to tell apart. According to a study published in the journal PNAS, in many cases it is impossible. Some people even find virtual faces more realistic than those of humans themselves.
The study mentions that the so-called “uncanny valley” or “mysterious valley” effect has been useful for users when it comes to checking if the person in a video or image is actually a Deepfake. This effect consists of the perception of poorly processed or inaccurate features, as well as unrealistic movements that these avatars created by artificial intelligence have. The problem is that technology has advanced so much that Deepfakes manage to circumvent this “mysterious valley”.
To verify this, a team of researchers created a series of digital faces through an automatic learning system. Specifically, they used a system of two neural networks that work together. On the one hand, a generator is responsible for creating the avatars starting from a “random matrix of pixels” and composing a final face that must then be examined by the second system, the discriminator. This checks the face created by artificial intelligence with a real face. If it finds differences, it penalizes the generator. Thus, until the system manages to create faces so exact that the discriminator cannot differentiate from the real ones. They managed to virtually replicate up to 400 faces of different races, ages and genders.
Deepfakes are not only indistinguishable, for some they are also more reliable than real faces
Real (R) and synthesized (S) faces with the confidence score received by the participants. Subsequently, they formed three groups with different participants. All of them had to distinguish the virtual faces from the real ones, but with different methods. The first group, made up of 315 people, compared only 182 of the 800 faces and, on average, only achieved 48.2 percent accuracy. The second group, of 219 participants, performed the same task: checking and distinguishing 182 faces. This time, instead, they enlisted the help of researchers to detect anomalies that could reveal whether or not a face was a deepfake. The results were better than those of the first 315 people, but it was not a noticeable change. On average, they were 59 percent correct.
Not only are synthetic faces highly realistic, but they are considered more reliable than real ones.
Hany Farid, co-author of the study. The last group, made up of 223 forgives, rated the confidence of these 128 images on a scale of one to seven. Where 1 is an image that is unreliable and therefore created by intelligence, and 7 is an image that is reliable and thus real. The results showed that the average score for real faces was 4.48, a score lower than the score for synthetic (fake) faces, which scored an average of 4.82 points.
Although these results clearly show the advances that have been achieved thanks to the latest technologies related to machine learning and artificial intelligence, they demonstrate the danger that the fact that a Deepfake is practically impossible to distinguish from a real one. It is not necessary to have advanced knowledge to use tools that can duplicate a person's face and, for example, impersonate a person.
